2026-10-01
M1/M2 receptor maps in a 68-region cortical model beat homogeneous networks on synchrony and information flow; local sleep-like slow waves emerge in awake-like states.
Whole-brain models have long shared a quiet assumption: every cortical region runs identical local dynamics, and whatever makes regions different is delegated to structural connectivity (SC), the white-matter matrix measured with diffusion MRI tractography. Real cortex violates this everywhere: receptor density, gene expression, and myelination all vary as gradients across the cortical sheet. Adding region-specific features such as the T1w:T2w ratio or excitation-inhibition balance does improve a model's fit to empirical functional connectivity, but whether molecular-scale heterogeneity shapes large-scale dynamics, and whether that depends on brain state, had no systematic answer.
Acetylcholine (ACh) is the natural entry point. It regulates both cortical excitability and neuronal adaptation; levels are high in wakefulness and REM sleep, low in deep sleep. The cholinergic projections reaching cortex from the basal forebrain are spatially nonuniform, and in mice the weakly innervated regions are exactly where local slow waves appear during REM sleep. A team at IDIBAPS Barcelona (Maria V. Sanchez-Vives's group) with Alain Destexhe at Paris-Saclay turned that observation into a computable question.
Three layers. Local dynamics: each region is an adaptive exponential mean-field model (AdEx-MF) coupling an excitatory and an inhibitory population plus an adaptation variable W, which stands for activity-dependent potassium currents and similar processes. Low adaptation gives irregular, desynchronized firing, awake-like; high adaptation gives Up/Down slow waves, sleep-like. That single bifurcation is the state knob of the whole framework. Connectivity: DTI tractography over 68 Desikan-Killiany regions, weights from fiber density, delays from tract length, interregional links excitatory, simulated in The Virtual Brain. Heterogeneity is the new ingredient. Expression of M1 (CHRM1) and M2 (CHRM2) muscarinic receptors, the two most abundant subtypes in human cortex, comes from the Allen Human Brain Atlas (spatial correlation between the two maps r = 0.526, Pspin = 0.0014); their average gives each region a density value (mean 1.03 ± 0.20) that modulates the local adaptation parameter after inversion: denser receptors mean weaker effective adaptation and higher excitability, since muscarinic activation suppresses the potassium currents behind adaptation. A factor hf slides from 0 (homogeneous) to 1 (fully heterogeneous) for a dose-response curve.
Robustness gets two checks: a PET-derived M1 receptor map reruns the entire analysis as an independent molecular modality, and three null models test specificity (spatially shuffled maps, Gaussian maps matched for mean and variance, spin-rotated surrogates preserving spatial autocorrelation; 100 realizations each).
The model first reproduces the state continuum: at mean adaptation 12 pA, low-amplitude 10 Hz activity with weak interregional correlations; at 24 pA, locally synchronized clusters and a 6 Hz peak; at 36 pA, network-wide 1 Hz Up/Down waves traveling across the cortex, with functional connectivity tightening onto structural connectivity just as empirical data show across states.
The payoff concentrates in information flow. Stimulating the superior parietal lobule for 30 ms and measuring transfer entropy (TE) to the other 67 regions over the following 400 ms:
| Network | TE, awake-like (10 pA) | TE, sleep-like (50 pA) |
| Transcriptomic heterogeneous | 1.38 ± 0.09 | 0.30 ± 0.07 |
| PET heterogeneous | 1.28 ± 0.08 | 0.28 ± 0.09 |
| Homogeneous | 1.23 ± 0.08 | 0.25 ± 0.06 |
On synchrony, mean FC overtakes the homogeneous network above 30 pA adaptation, and the phase lag index (PLI, insensitive to zero-lag correlations, suited to traveling waves) is higher across the entire range, rising monotonically with hf. Specificity is mixed: six of twelve null comparisons reach P < 0.05, four of them in the sleep-like regime (z = 2.36-2.89 against shuffled maps in the sleep-like state; z = 3.32 against spin-rotated maps at intermediate adaptation with hf = 1), while Gaussian maps show no significant difference at intermediate adaptation (z no higher than 1.52). Awake-state TE is also statistically indistinguishable from the Gaussian null. The authors' reading: generic heterogeneity already preserves broad state transitions, and biologically structured heterogeneity adds a selective boost within synchronized states.
The intermediate regime (24 pA) produces the paper's most striking phenomenon: a few regions fall into Up/Down slow waves while the rest stay awake-like, two dynamics coexisting in one network. Which regions sleep first is decided jointly by two only weakly correlated factors (r = -0.23, P = 0.056): regional adaptation level correlates positively with delta power (0-4 Hz), and in-degree connectivity correlates negatively, so regions receiving less excitatory input drift toward sleep first. Under the null maps the adaptation-delta correlation collapses and the in-degree-delta correlation falls back toward the homogeneous baseline: the receptor map's spatial organization amplifies adaptation's role and loosens the connectome's sole claim on where slow waves land. This lines up with the mouse finding that weakly innervated regions show local slow waves first.
For AI practitioners this is corroborating evidence from the biological side. Spiking-network and oscillator studies have long argued that heterogeneous units code better and synchronize better (the cited "Random heterogeneity outperforms design in network synchronization" is representative); this work pushes the principle to a biologically constrained human-scale network, where the brain does not average heterogeneity away but organizes it with molecular gradients into an amplifier for cortex-wide communication. Anyone designing neuromorphic hardware or heterogeneous-unit networks gains a design argument, though not a ready-to-use recipe.
For neuroscience, the framework puts receptor maps (transcriptomic plus PET) inside a whole-brain model that also generates state transitions, and delivers a mechanistic account of local sleep: regions with high adaptation and low in-degree fall asleep first. Attentional lapses, performance drops under sleep deprivation, and slow waves around brain lesions all fit the same frame, and the authors point at personalized lesion models as the next step.
The authors list three: cholinergic action is reduced to a single adaptation pathway while muscarinic receptors act on many synaptic processes; M1 and M2 are merged into one parameter, and the PET arm uses M1 only; the maps are static population averages with no temporal dynamics. The null results also cap the claim: broad state transitions do not depend on the receptor map's precise spatial organization, so heterogeneity acts as a modulator.
Two more caveats from a close read: absolute effect sizes are small, with PLI differences on the order of 0.001 to 0.015 while the paper reports standardized z scores; and the local-slow-wave account is validated only by correlations inside the model, not against empirical EEG or intracranial maps of where slow waves actually occur. The account is self-consistent, not yet pinned by external data.